1. Special Topics BSC4933/5936 Florida State University The Department of Biological Science http://www.bio.fsu.edu http://www.bio.fsu.edu Sept. 23, 2003 An Introduction to Bioinformatics

2. How can you search the databases for similar sequences, if pair-wise alignments take N 2 time?! Significance and heuristics... Database Similarity Searching Steven M. Thompson Steven M. Thompson Florida State University School of Computational Science and Information Technology (CSIT) CSIT

3. But, why even do database searches? We can imagine screening databases for sequences similar to ours using the concepts of dynamic programming and log-odds scoring matrices and some yet to be described tricks. But what do database searches tell us; what can we gain from them? Why even bother? Inference through homology is a fundamental principle of biology. When a sequence is found to fall into a preexisting family we may be able to infer function, mechanism, evolution, perhaps even structure, based on homology with its neighbors. If no significant similarity can be found, the very fact that your sequence is new and different could be very important. Granted, its characterization may prove difficult, but it could be well worth it.

4. Homology and similarity — Don’t confuse homology with similarity: there is a huge difference! Similarity is a statistic that describes how much two (sub)sequences are alike according to some set scoring criteria. It can be normalized to ascertain statistical significance, but it’s still just a number. Homology, in contrast and by definition, implies an evolutionary relationship — more than just the fact that we’ve all evolved from the same old primordial ‘ooze.’ To demonstrate homology reconstruct the phylogeny of the organisms or genes of interest. Better yet, show experimental evidence — structural, morphological, genetic, or fossil — that corroborates your assertion. There is no such thing as percent homology; something is either homologous or it is not. Walter Fitch is credited with “homology is like pregnancy — you can’t be 45% pregnant, just like something can’t be 45% homologous. You either are or you are not.” Highly significant similarity can argue for homology, but never the other way around.

5. So, first — Significance : when is any alignment worth anything biologically? An old statistics trick — Monte Carlo simulations: Z score = [ ( actual score ) - ( mean of randomized scores ) ] ( standard deviation of randomized score distribution ) ( standard deviation of randomized score distribution ) Independent of all that, what is a ‘good’ alignment?

6. The Normal (Abby Normal?) distribution — Many Z scores measure the distance from the mean using this simplistic Monte Carlo model assuming a Gaussian distribution, aka the Normal distribution (http://mathworld.wolfram.com/NormalDistribution.html), Normal distributionhttp://mathworld.wolfram.com/NormalDistribution.htmlNormal distributionhttp://mathworld.wolfram.com/NormalDistribution.html in spite of the fact that ‘sequence-space’ actually follows what is know as the ‘Extreme Value distribution.’ However, the Monte Carlo method does approximate significance estimates fairly well.

7. < 20 650 0:== 22 0 0: 22 0 0: 24 3 0:= 24 3 0:= 26 22 8:* 26 22 8:* 28 98 87:* 28 98 87:* 30 289 528:* 30 289 528:* 32 1714 2042:===* 32 1714 2042:===* 34 5585 5539:=========* 34 5585 5539:=========* 36 12495 11375:==================*== 36 12495 11375:==================*== 38 21957 18799:===============================*===== 38 21957 18799:===============================*===== 40 28875 26223:===========================================*==== 40 28875 26223:===========================================*==== 42 34153 32054:=====================================================*=== 42 34153 32054:=====================================================*=== 44 35427 35359:==========================================================* 44 35427 35359:==========================================================* 46 36219 36014:===========================================================* 46 36219 36014:===========================================================* 48 33699 34479:======================================================== * 48 33699 34479:======================================================== * 50 30727 31462:=================================================== * 50 30727 31462:=================================================== * 52 27288 27661:=============================================* 52 27288 27661:=============================================* 54 22538 23627:====================================== * 54 22538 23627:====================================== * 56 18055 19736:============================== * 56 18055 19736:============================== * 58 14617 16203:========================= * 58 14617 16203:========================= * 60 12595 13125:=====================* 60 12595 13125:=====================* 62 10563 10522:=================* 62 10563 10522:=================* 64 8626 8368:=============*= 64 8626 8368:=============*= 66 6426 6614:==========* 66 6426 6614:==========* 68 4770 5203:========* 68 4770 5203:========* 70 4017 4077:======* 70 4017 4077:======* 72 2920 3186:=====* 72 2920 3186:=====* 74 2448 2484:====* 74 2448 2484:====* 76 1696 1933:===* 76 1696 1933:===* 78 1178 1503:==* 78 1178 1503:==* 80 935 1167:=* 80 935 1167:=* 82 722 893:=* 82 722 893:=* 84 454 707:=* 84 454 707:=* 86 438 547:* 86 438 547:* 88 322 423:* 88 322 423:* 90 257 328:* 90 257 328:* 92 175 253:* 92 175 253:* 94 210 196:* 94 210 196:* 96 102 152:* 96 102 152:* 98 63 117:* 98 63 117:* 100 58 91:* 100 58 91:* 102 40 70:* 102 40 70:* 104 30 54:* 104 30 54:* 106 17 42:* 106 17 42:* 108 14 33:* 108 14 33:* 110 14 25:* 110 14 25:* 112 12 20:* 112 12 20:* 114 9 15:* 114 9 15:* 116 6 12:* 116 6 12:* 118 8 9:* 118 8 9:* >120 1030 7:*= Based on this known statistical distribution, and robust statistical methodology, a realistic Expectation function, the E Value, can be calculated from database searches. The particulars of how this is done will wait, but the ‘take-home’ message is the same... ‘Sequence-space’ (Huh, what’s that?) actually follows the ‘Extreme Value distribution’ Extreme Value distributionExtreme Value distribution (http://mathworld.wolfram.com/ExtremeValueDistribution.html). http://mathworld.wolfram.com/ExtremeValueDistribution.html

8. The Expectation Value? The higher the E value is, the more probable that the observed match is due to chance in a search of the same size database, and the lower its Z score will be, i.e. is NOT significant. Therefore, the smaller the E value, i.e. the closer it is to zero, the more significant it is and the higher its Z score will be! The E value is the number that really matters.

9. Rules of thumb for a protein search — The Z score represents the number of standard deviations some particular alignment is from a distribution of random alignments (the normal distribution). They very roughly correspond to the listed E Values (based on the Extreme Value distribution) for a typical protein sequence similarity search through a database with ~125,000 protein entries.

10. On to the searches — But N 2 is way too slow! How can it be done? Database searching programs use the two concepts of dynamic programming and log-odds scoring matrices; however, dynamic programming takes far too long when used against most sequence databases with a ‘normal’ computer. Remember how big the databases are! Therefore, the programs use tricks to make things happen faster. These tricks fall into two main categories, that of hashing, and that of approximation.

11. Corn beef hash? Huh... Hashing is the process of breaking your sequence into small ‘words’ or ‘k-tuples’ (think all chopped up, just like corn beef hash) of a set size and creating a ‘look-up’ table with those words keyed to position numbers. Computers can deal with numbers way faster than they can deal with strings of letters, and this preprocessing step happens very quickly. Then when any of the word positions match part of an entry in the database, that match, the ‘offset,’ is saved. In general, hashing reduces the complexity of the search problem from N 2 for dynamic programming to N, the length of all the sequences in the database.

12. A simple hash table — (this example is from the Krane and Raymer text p.50) The sequence FAMLGFIKYLPGCM and a word size of one, would produce this query lookup hash table: wordACFGIKLMPY Pos.213157843119 6121014 comparing it to the database sequence TGFIKYLPGACT, would yield the following offset table: 1 23456789101112 TGFIKYLPGACTTGFIKYLPGACTTGFIKYLPGACTTGFIKYLPGACT 3-2333-33-4-82 10333

13. Hmmm & some interpretation — The offset numbers come from the difference between the positions of the words in the query sequence and the position of the occurrence of that word in the target sequence. Then.... Look at all of the offsets equal to three in the previous table. Therefore, offset the alignment by three: FAMLGFIKYLPGCM |||||||| |||||||| TGFIKYLPGACT TGFIKYLPGACT Quick and easy. Computers can compare these sorts of tables very fast. The trick is to ‘know’ how far to attempt to extend the alignment out.

14. OK. Heuristics... What’s that? Approximation techniques are collectively known as ‘heuristics.’ Webster’s defines heuristic as “serving to guide, discover, or reveal;... but unproved or incapable of proof.” In database similarity searching techniques the heuristic usually restricts the necessary search space by calculating some sort of a statistic that allows the program to decide whether further scrutiny of a particular match should be pursued. This statistic may miss things depending on the parameters set — that’s what makes it heuristic. ‘Worthwhile’ results at the end are compiled and the longest alignment within the program’s restrictions is created. The exact implementation varies between the different programs, but the basic idea follows in most all of them.

15. Two predominant versions exist: BLAST and Fast Both return local alignments, and are not a single program, but rather a family of programs with implementations designed to compare a sequence to a database in about every which way imaginable. These include: 1)a DNA sequence against a DNA database (not recommended unless forced to do so because you are dealing with a non-translated region of the genome — DNA is just too darn noisy, only identity & four bases!), 2)a translated (where the translation is done ‘on-the-fly’ in all six frames) version of a DNA sequence against a translated (‘on-the-fly’ six-frame) version of the DNA database (not available in the Fast package), 3)a translated (‘on-the-fly’ six-frame) version of a DNA sequence against a protein database, 4)a protein sequence against a translated (‘on-the-fly’ six-frame) version of a DNA database, 5)or a protein sequence against a protein database. Many implementations allow for the possibility of frame shifts in translated comparisons and don’t penalize the score for doing so.

16. The BLAST and Fast programs — some generalities BLAST — Basic Local Alignment Search Tool, developed at NCBI. 1)Normally NOT a good idea to use for DNA against DNA searches w/o translation (not optimized); 2)Pre-filters repeat and “low complexity” sequence regions; 4)Can find more than one region of gapped similarity; 5)Very fast heuristic and parallel implementation; 6)Restricted to precompiled, specially formatted databases; FastA — and its family of relatives, developed by Bill Pearson at the University of Virginia. 1)Works well for DNA against DNA searches (within limits of possible sensitivity); 2)Can find only one gapped region of similarity; 3)Relatively slow, should often be run in the background; 4)Does not require specially prepared, preformatted databases.

17. The algorithms, in brief — BLAST: Fast: Two word hits on the same diagonal above some similarity threshold triggers ungapped extension until the score isn’t improved enough above another threshold: the HSP. Find all ungapped exact word hits; maximize the ten best continuous regions’ scores: init1. Combine non- overlapping init regions on different diagonals: initn. Use dynamic programming ‘in a band’ for all regions with initn scores better than some threshold: opt score. Initiate gapped extensions using dynamic programming for those HSP’s above a third threshold up to the point where the score starts to drop below a fourth threshold: yields alignment.

18. BLAST — the algorithm in more detail 1)After BLAST has sorted its lookup table, it tries to find all double word hits along the same diagonal within some specified distance using what NCBI calls a Discrete Finite Automaton (DFA). These word hits of size W do not have to be identical; rather, they have to be better than some threshold value T. To identify these double word hits, the DFA scans through all strings of words (typically W=3 for peptides) that score at least T (usually 11 for peptides). 2)Each double word hit that passes this step then triggers a process called un-gapped extension in both directions, such that each diagonal is extended as far as it can, until the running score starts to drop below a pre-defined value X within a certain range A. The result of this pass is called a High-Scoring segment Pair or HSP. 3)Those HSPs that pass this step with a score better than S then begin a gapped extension step utilizing dynamic programming. Those gapped alignments with Expectation values better than the user specified cutoff are reported. The extreme value distribution of BLAST Expectation values is pre-computed against each precompiled database — this is one area that speeds up the algorithm considerably.

19. The BLAST algorithm, continued The math can be generalized thus: for any two sequences of length m and n, local, best alignments are identified as HSPs. HSPs are stretches of sequence pairs that cannot be further improved by extension or trimming, as described above. For un-gapped alignments, the number of expected HSPs with a score of at least S is given by the formula: E = Kmne  s This is called an E-value for the score S. In a database search n is the size of the database in residues, so N=mn is the search space size. K and are be supplied by statistical theory, and, as mentioned above, can be calculated by comparison to pre-computed, simulated distributions. These two parameters define the statistical significance of an E-value. The E-value defines the significance of the search. As mentioned above, the smaller an E-value is, the more likely it is significant. A value of 0.01 to 0.001 is a good starting point for significance in most typical searches. In other words, in order to assess whether a given alignment constitutes evidence for homology, it helps to know how strong an alignment can be expected from chance alone.

20. The Fast algorithm — in more detail Fast is an older algorithm than BLAST. The original Fast paper came out in 1988, based on David Lipman’s work in a 1983 paper; the original BLAST paper was published in 1990. Both algorithms have been upgraded substantially since originally released. Fast was the first widely used, powerful sequence database searching algorithm. Bill Pearson continually refines the programs such that they remain a viable alternative to BLAST, especially if one is restricted to searching DNA against DNA without translation. They are also very helpful in situations where BLAST finds no significant alignments — arguably, Fast may be more sensitive than BLAST in these situations. Fast is also a hashing style algorithm and builds words of a set k- tuple size, by default two for peptides. It then identifies all exact word matches between the sequence and the database members. Note that the word matches must be exact for Fast and only similar, above some threshold, for BLAST.

21. The Fast algorithm, continued From these exact word matches: 1)Scores are assigned to each continuous, ungapped, diagonal by adding all of the exact match BLOSUM values. 2)The ten highest scoring diagonals for each query-database pair are then rescored using BLOSUM similarities as well as identities and ends are trimmed to maximize the score. The best of each of these is called the Init1 score. 3)Next the program ‘looks’ around to see if nearby off-diagonal Init1 alignments can be combined by incorporating gaps. If so, a new score, Initn, is calculated by summing up all the contributing Init1 scores, penalizing gaps with a penalty for each. 4)The program then constructs an optimal local alignment for all Initn pairs with scores better than some set threshold using a variation of dynamic programming “in a band.” A sixteen residue band centered at the highest Init1 region is used by default with peptides. The score generated from this step called opt.

22. The Fast algorithm, still continued 5)Next, Fast uses a simple linear regression against the natural log of the search set sequence length to calculate a normalized z-score for the sequence pair. Note that this is not the same Monte Carlo style Z score described earlier, and can not be directly compared to one. 6)Finally, it compares the distribution of these z-scores to the actual extreme-value distribution of the search. Using this distribution, the program estimates the number of sequences that would be expected to have, purely by chance, a z-score greater than or equal to the z-score obtained in the search. This is reported as the Expectation value. 7)If the user requests pair-wise alignments in the output, then the program uses full Smith-Waterman local dynamic programming, not ‘restricted to a band,’ to produce its final alignments.

23. Let’s see ‘em in action To begin we’ll go to the most widely used (and abused!) biocomputing program on earth: NCBI’s BLAST — Connect to NCBI’s BLAST page with any Web browser: http://www.ncbi.nlm.nih.gov/BLAST/http://www.ncbi.nlm.nih.gov/BLAST/. http://www.ncbi.nlm.nih.gov/BLAST/ There is a wealth of information there, including a wonderful tutorial and several very good essays for teaching yourself way more about BLAST than this lecture can ever hope for. For now I’ll demonstrate with a simple example, one of my favorites, the elongation factor 1  protein from Giardia lamblia, named EF1A_Giala in the Swiss-Prot database, but we have to use the accession code, Q08046, for NCBI’s BLAST server to find the sequence. Let’s see how it works and how quickly we get results back.

24. Let’s contrast that with GCG’s BLAST version. I’ll illustrate with the same molecule and I’ll use GCG’s SeqLab GUI to show the difference between the two implementations of the program. SeqLab And finally, let’s see how GCG’s FastA version compares to either BLAST implementation. Again, I’ll launch the program from SeqLab with the same example, but this time I’ll take advantage of Fast’s flexible database search syntax, being able to use any valid GCG sequence specification. Here I’ll search against a precompiled LookUp list file of all of the so-called ‘primitive’ eukaryotes in Swiss-Prot.

25. Finally — Why do I keep ‘diss’ing’ DNA for searches and alignment? All database similarity searching and sequence alignment, regardless of the algorithm used, is far more sensitive at the amino acid level than at the DNA level. This is because proteins have twenty match criteria versus DNA’s four, and those four DNA bases can generally only be identical, not similar, to each other; and many DNA base changes (especially third position changes) do not change the encoded protein. All of these factors drastically increase the ‘noise’ level of a DNA against DNA search, and give protein searches a much greater ‘look-back’ time, typically doubling it. Therefore, whenever dealing with coding sequence, it is always prudent to search at the protein level!

26. Altschul, S. F., Gish, W., Miller, W., Myers, E. W., and Lipman, D. J. (1990) Basic Local Alignment Tool. Journal of Molecular Biology 215, 403-410. Altschul, S.F., Madden, T.L., Schaffer, A.A., Zhang, J., Zhang, Z., Miller, W., and Lipman, D.J. (1997) Gapped BLAST and PSI-BLAST: a New Generation of Protein Database Search Programs. Nucleic Acids Research 25, 3389-3402. Genetics Computer Group (GCG) (Copyright 1982-2002) Program Manual for the Wisconsin Package, Version 10.3, Accelrys, Inc. A Pharmocopeia Company, San Diego, California, U.S.A. Gribskov, M. and Devereux, J., editors (1992) Sequence Analysis Primer. W.H. Freeman and Company, New York, New York, U.S.A. Henikoff, S. and Henikoff, J.G. (1992) Amino Acid Substitution Matrices from Protein Blocks. Proceedings of the National Academy of Sciences U.S.A. 89, 10915-10919. Needleman, S.B. and Wunsch, C.D. (1970) A General Method Applicable to the Search for Similarities in the Amino Acid Sequence of Two Proteins. Journal of Molecular Biology 48, 443-453. Pearson, W.R. and Lipman, D.J. (1988) Improved Tools for Biological Sequence Analysis. Proceedings of the National Academy of Sciences U.S.A. 85, 2444-2448. Schwartz, R.M. and Dayhoff, M.O. (1979) Matrices for Detecting Distant Relationships. In Atlas of Protein Sequences and Structure, (M.O. Dayhoff editor) 5, Suppl. 3, 353-358, National Biomedical Research Foundation, Washington D.C., U.S.A. Smith, T.F. and Waterman, M.S. (1981) Comparison of Bio-Sequences. Advances in Applied Mathematics 2, 482-489. Wilbur, W.J. and Lipman, D.J. (1983) Rapid Similarity Searches of Nucleic Acid and Protein Data Banks. Proceedings of the National Academy of Sciences U.S.A. 80, 726-730. References and a Comment: The better you understand the chemical, physical, and biological systems involved, the better your chance of success in analyzing them. Certain strategies are inherently more appropriate to others in certain circumstances. Making these types of subjective, discriminatory decisions and utilizing all of the available options so that you can generate the most practical data for evaluation are two of the most important ‘take-home’ messages that I can offer!